1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2009 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
20 /* Various options to tweak for performance benchmarking purposes. */
22 /* Perform all get/put operations synchronously. */
23 int synchronous_stores;
25 /* Write data in cache immediately after file is modified. */
26 int writethrough_cache;
28 /* Can inodes be fetched asynchronously? (Inode object is initially
29 * created in a pending state, and not unlocked until the data is actually
31 int sync_inode_fetches;
33 /* Should frontends handle requests serially or allow operations to proceed
37 /* Target size of the disk cache at the proxy, in kilobytes. */
41 extern BlueSkyOptions bluesky_options;
43 /* Maximum number of threads to use in any particular thread pool, or -1 for no
45 extern int bluesky_max_threads;
47 /* A general-purpose counter for gathering run-time statistics. */
48 struct bluesky_stats {
53 struct bluesky_stats *bluesky_stats_new(const char *name);
54 void bluesky_stats_add(struct bluesky_stats *stats, int64_t value);
55 void bluesky_stats_dump_all();
57 /* BlueSky status and error codes. Various frontends should translate these to
58 * the appropriate error code for whatever protocol they implement. */
60 BSTATUS_OK = 0, /* No error */
61 BSTATUS_IOERR, /* I/O error of some form */
62 BSTATUS_NOENT, /* File does not exist */
65 void bluesky_init(void);
67 gchar *bluesky_lowercase(const gchar *s);
69 struct _BlueSkyCacheFile;
70 typedef struct _BlueSkyCacheFile BlueSkyCacheFile;
74 BlueSkyCacheFile *mmap;
79 BlueSkyRCStr *bluesky_string_new(gpointer data, gsize len);
80 BlueSkyRCStr *bluesky_string_new_from_gstring(GString *s);
81 BlueSkyRCStr *bluesky_string_new_from_mmap(BlueSkyCacheFile *mmap,
82 int offset, gsize len);
83 void bluesky_string_ref(BlueSkyRCStr *string);
84 void bluesky_string_unref(BlueSkyRCStr *string);
85 BlueSkyRCStr *bluesky_string_dup(BlueSkyRCStr *string);
86 void bluesky_string_resize(BlueSkyRCStr *string, gsize len);
88 /* Cryptographic operations. */
89 #define CRYPTO_BLOCK_SIZE 16 /* 128-bit AES */
90 #define CRYPTO_KEY_SIZE 16
91 #define CRYPTO_HASH_SIZE 32 /* SHA-256 */
93 struct BlueSkyCryptKeys {
94 uint8_t encryption_key[CRYPTO_KEY_SIZE];
95 uint8_t authentication_key[CRYPTO_HASH_SIZE];
98 void bluesky_crypt_init();
99 void bluesky_crypt_hash_key(const char *keystr, uint8_t *out);
100 void bluesky_crypt_random_bytes(guchar *buf, gint len);
101 BlueSkyRCStr *bluesky_crypt_encrypt(BlueSkyRCStr *in, const uint8_t *key);
102 BlueSkyRCStr *bluesky_crypt_decrypt(BlueSkyRCStr *in, const uint8_t *key);
104 /* Storage interface. This presents a key-value store abstraction, and can
105 * have multiple implementations: in-memory, on-disk, in-cloud. */
106 struct _BlueSkyStore;
107 typedef struct _BlueSkyStore BlueSkyStore;
110 typedef struct _BlueSkyLog BlueSkyLog;
112 struct _BlueSkyCloudLogState;
113 typedef struct _BlueSkyCloudLogState BlueSkyCloudLogState;
115 struct _BlueSkyCloudLog;
116 typedef struct _BlueSkyCloudLog BlueSkyCloudLog;
118 void bluesky_store_init();
119 BlueSkyStore *bluesky_store_new(const gchar *type);
120 void bluesky_store_free(BlueSkyStore *store);
121 BlueSkyRCStr *bluesky_store_get(BlueSkyStore *store, const gchar *key);
122 void bluesky_store_put(BlueSkyStore *store,
123 const gchar *key, BlueSkyRCStr *val);
125 /* File types. The numeric values are chosen to match with those used in
129 BLUESKY_DIRECTORY = 2,
131 BLUESKY_CHARACTER = 4,
136 /* Special types used only internally. */
137 BLUESKY_PENDING = 0, /* Inode being loaded; type not yet determined */
138 BLUESKY_INVALID = -1, /* Inode is invalid (failed to load) */
141 /* Filesystem state. Each filesystem which is exported is represented by a
142 * single bluesky_fs structure in memory. */
146 gchar *name; /* Descriptive name for the filesystem */
147 GHashTable *inodes; /* Cached inodes */
148 uint64_t next_inum; /* Next available inode for allocation */
152 BlueSkyCloudLogState *log_state;
154 /* Accounting for memory used for caches. Space is measured in blocks, not
155 * bytes. Updates to these variables must be made atomically. */
158 /* Like above, but tracking data stored in the cloudlog entries
160 * - cache_log_dirty: data uncommitted to journal and cloud
161 * - cache_log_writeback: data being written to journal
162 * - cache_log_journal: data committed to journal
163 * - cache_log_cloud: data written to cloud as well
164 * Log entries should progress from the top state to the bottom, and are
165 * only ever counted in one category at a time. */
166 gint cache_log_dirty, cache_log_writeback,
167 cache_log_journal, cache_log_cloud;
169 /* Linked list of inodes, sorted by access/modification times for cache
170 * management. Editing these lists is protected by the filesystem lock; to
171 * avoid deadlock do not attempt to take any other locks while the FS lock
172 * is held for list editing purposes. Items at the head of the list are
173 * most recently accessed/modified. */
174 GList unlogged_list; // Changes not yet synced to journal
175 GList dirty_list; // Not yet written to cloud storage
176 GList accessed_list; // All in-memory inodes
178 /* Mutex for the flush daemon, to prevent concurrent execution. */
181 /* Used to wait for the cache daemon to free up space */
184 /* Mapping of object identifiers (blocks, inodes) to physical location (in
185 * the local cache or in the logs in the cloud). */
186 GHashTable *locations;
188 /* The inode map, which maps inode numbers to the location of the most
190 GSequence *inode_map;
192 /* Queue for asynchronous cloudlog unrefs, where needed. */
193 GAsyncQueue *unref_queue;
195 /* Thread pool for asynchronous inode fetches */
196 GThreadPool *inode_fetch_thread_pool;
199 /* Inode number of the root directory. */
200 #define BLUESKY_ROOT_INUM 1
202 /* Timestamp, measured in microseconds since the Unix epoch. */
203 typedef int64_t bluesky_time;
205 /* High-resolution timer, measured in nanoseconds. */
206 typedef int64_t bluesky_time_hires;
207 bluesky_time_hires bluesky_now_hires();
209 /* In-memory representation of an inode within a Blue Sky server. This
210 * corresponds roughly with information that is committed to persistent
211 * storage. Locking/refcounting rules:
212 * - To access or modify any data fields, the lock must be held. This
213 * includes file blocks.
214 * - One reference is held by the BlueSkyFS inode hash table. If that is the
215 * only reference (and the inode is unlocked), the inode is subject to
216 * dropping from the cache.
217 * - Any pending operations should hold extra references to the inode as
218 * appropriate to keep it available until the operation completes.
219 * - Locking dependency order is, when multiple locks are to be acquired, to
220 * acquire locks on parents in the filesystem tree before children.
221 * (TODO: What about rename when we acquire locks in unrelated parts of the
223 * - An inode should not be locked while the filesystem lock is already held,
224 * since some code may do an inode lookup (which acquires the filesystem
225 * lock) while a different inode is locked.
233 BlueSkyFileType type;
238 /* Rather than track an inode number and generation number, we will simply
239 * never re-use a fileid after a file is deleted. 64 bits should be enough
240 * that we don't exhaust the identifier space. */
243 /* change_count is increased with every operation which modifies the inode,
244 * and can be used to determine if cached data is still valid.
245 * change_commit is the value of change_count when the inode was last
246 * committed to stable storage (the log).
247 * change_cloud tracks which version was last commited to cloud storage. */
248 uint64_t change_count, change_commit, change_cloud;
250 /* Timestamp for controlling when modified data is flushed to stable
251 * storage. When an inode is first modified from a clean state, this is
252 * set to the current time. If the inode is clean, it is set to zero. */
255 /* Last access time to this inode, for controlling cache evictions. */
258 /* Version of the object last serialized and committed to storage. */
259 BlueSkyCloudLog *committed_item;
261 /* Pointers to the linked-list elements for this inode in the accessed and
262 * dirty linked lists. We re-use the GList structure, using ->next to
263 * point to the head of the list and ->prev to point to the tail. The data
264 * element is unused. */
265 GList *unlogged_list, *accessed_list, *dirty_list;
267 int64_t atime; /* Microseconds since the Unix epoch */
270 int64_t ntime; /* "new time": time object was created */
272 /* File-specific fields */
276 /* Directory-specific fields */
277 GSequence *dirents; /* List of entries for READDIR */
278 GHashTable *dirhash; /* Hash table by name for LOOKUP */
279 GHashTable *dirhash_folded; /* As above, but case-folded */
280 uint64_t parent_inum; /* inode for ".."; 0 if the root directory */
282 /* Symlink-specific fields */
283 gchar *symlink_contents;
286 /* A directory entry. The name is UTF-8 and is a freshly-allocated string.
287 * Each directory entry is listed in two indices: dirents is indexed by cookie
288 * and dirhash by name. The cookie is a randomly-assigned 32-bit value, unique
289 * within the directory, that remains unchanged until the entry is deleted. It
290 * is used to provide a stable key for restarting a READDIR call. */
293 gchar *name_folded; /* Name, folded for case-insensitive lookup */
298 /* File data is divided into fixed-size blocks (except the last block which may
299 * be short?). These blocks are backed by storage in a key/value store, but
300 * may also be dirty if modifications have been made in-memory that have not
302 #define BLUESKY_BLOCK_SIZE 32768ULL
303 #define BLUESKY_MAX_FILE_SIZE (BLUESKY_BLOCK_SIZE << 24)
305 BLUESKY_BLOCK_ZERO = 0, /* Data is all zeroes, not explicitly stored */
306 BLUESKY_BLOCK_REF = 1, /* Reference to cloud log item, data clean */
307 BLUESKY_BLOCK_DIRTY = 2, /* Data needs to be committed to store */
311 BlueSkyBlockType type;
312 BlueSkyCloudLog *ref; /* if REF: cloud log entry with data */
313 BlueSkyRCStr *dirty; /* if DIRTY: raw data in memory */
316 BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store);
318 gboolean bluesky_inode_is_ready(BlueSkyInode *inode);
320 int64_t bluesky_get_current_time();
321 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime);
322 uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs);
323 void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type);
324 BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs, BlueSkyFileType type);
326 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum);
327 void bluesky_inode_ref(BlueSkyInode *inode);
328 void bluesky_inode_unref(BlueSkyInode *inode);
329 void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode);
331 void bluesky_dirent_destroy(gpointer dirent);
332 uint64_t bluesky_directory_lookup(BlueSkyInode *inode, gchar *name);
333 uint64_t bluesky_directory_ilookup(BlueSkyInode *inode, gchar *name);
334 BlueSkyDirent *bluesky_directory_read(BlueSkyInode *dir, uint32_t cookie);
335 gboolean bluesky_directory_insert(BlueSkyInode *dir, const gchar *name,
337 void bluesky_directory_dump(BlueSkyInode *dir);
339 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size);
340 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
341 const char *data, gint len);
342 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
343 char *buf, gint len);
345 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum);
347 gint bluesky_dirent_compare(gconstpointer a, gconstpointer b,
350 void bluesky_flushd_invoke(BlueSkyFS *fs);
351 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs);
352 void bluesky_inode_do_sync(BlueSkyInode *inode);
353 void bluesky_flushd_thread_launch(BlueSkyFS *fs);
355 void bluesky_debug_dump(BlueSkyFS *fs);